The 13% traveled. The authors' caveat did not.

TL;DR. Brynjolfsson, Chandar and Chen used ADP payroll records covering millions of US workers and found a 13% relative employment decline for workers aged 22 to 25 in the most AI-exposed occupations, rising to 16% in a later version, with around 20% for young software developers since late 2022. The declines concentrate where AI automates rather than augments, and experienced workers are largely unaffected. That headline traveled everywhere. In a February 2026 update the authors themselves reported that, with the broadest set of controls, the decline in AI-exposed occupations only becomes significant in 2024, and that earlier declines were likely influenced by non-AI factors. That qualification did not travel. And a competing account attributes the same pattern to the sharpest monetary tightening in four decades, which predicts exactly this age gradient.

---

Status: real pattern, contested cause. Primary sources: the Stanford Digital Economy Lab working paper and its Canaries dashboard, including the authors' February 2026 update; and the Economic Innovation Group's January 2026 critique. This article does not resolve the causal question. It reports what each side established and what the authors said about their own result.

---

What was measured

ADP is the largest payroll provider in the United States. The study used its monthly individual-level records for millions of workers, running through September 2025, and linked them to established measures of occupational exposure to generative AI.

That is unusually good data for this question. Not a survey, not job postings, not announcements. Payroll.

The findings, as published:

Employment for workers aged 22 to 25 in the most AI-exposed occupations fell 13% relative to less-exposed occupations, after controlling for firm-level shocks. A later version reports 16%.

For software developers aged 22 to 25 specifically, headcount fell around 20% since late 2022.

The declines concentrate in occupations where AI automates tasks rather than augmenting them. Jobs described as augmented by AI did not show the same pattern.

And the effect is age-specific. The market for experienced workers held up; entry-level stagnated.

This is a real pattern in good data, and nothing below disputes that.

What the authors said next

In a February 2026 update, the authors reported that when the broadest set of controls is included, the timing of decline in AI-exposed occupations only becomes significant in 2024, and that earlier declines were likely influenced by non-AI factors.

Read that carefully. The generative AI moment is dated to late 2022. The headline figure runs from that point. The authors' own further analysis says the statistically significant portion begins in 2024, and that what happened before was probably something else.

That does not overturn the finding. It narrows the window in which the finding is attributable, which is what careful researchers do when they add controls and the picture changes.

It also did not travel. The 13% is quoted constantly. The February 2026 qualification appears in the dashboard documentation and almost nowhere else.

This is citation decay with an unusual property: the caveat came from the authors, was published, and still lost the race to their own headline. A number does not need decades or a broken link to shed its qualifications. It needs a wide gap in quotability.

The competing explanation

The Economic Innovation Group published a critique in January 2026 arguing the pattern is not early technological displacement but the predictable consequence of the sharpest monetary policy tightening cycle in four decades.

The mechanism is specific and it fits. Rate rises from 2022 collapsed hiring. Job postings fell sharply. When firms stop hiring, they stop hiring at the bottom first, because entry-level roles are the marginal ones and experienced staff are retained. The primary entry points to the labour market and the pathways for progression disappear, leaving young workers unable to get onto the ladder.

A disproportionate negative effect on 22 to 25 year olds is precisely what that theory predicts, without any reference to AI.

EIG states its interpretation joins a body of other studies reaching similar conclusions.

Why this is hard to settle

The two candidate causes happened at the same time. Generative AI reached wide adoption in late 2022. The tightening cycle ran from 2022. Any analysis has to separate two shocks that share a start date, which is close to the hardest identification problem in applied economics.

The strongest evidence for the AI account is the automation-versus-augmentation split. A monetary shock should hit entry-level hiring across exposed and unexposed occupations similarly; it has no obvious reason to distinguish jobs where AI automates from jobs where AI augments. That distinction is the AI hypothesis's best asset, and it is the finding most worth watching.

The strongest evidence for the macroeconomic account is that the age gradient is exactly what a hiring freeze produces, and that the authors' own broadest-control specification pushes significance to 2024, after the initial adoption wave.

Neither is dispositive, and this article does not pick one.

What can be said without picking

Entry-level employment in AI-exposed occupations declined. Both accounts agree.

Young workers are bearing the adjustment, whatever its cause. That is not in dispute and it matters to the people it is happening to regardless of which mechanism produced it.

The two explanations imply different responses. If it is monetary, the effect unwinds when hiring recovers. If it is structural automation, it does not, and the entry-level rung does not come back when rates fall. Watching what happens to entry-level hiring as monetary conditions ease is the natural test, and it is running now.

And the automation-augmentation split is the variable to track. If exposed-and-automated diverges further from exposed-and-augmented as the macroeconomic shock recedes, that is evidence the AI account was right.

Three things this establishes

Author-issued caveats do not automatically travel with findings. The qualification here was published by the same team, in the same project, and lost to its own headline. Anyone citing a working paper should check what its authors have said since, which is a cheap habit almost nobody has.

Simultaneous shocks are close to unidentifiable. Two large causes with the same start date, acting on the same population, cannot be separated by controls alone. Time is what will separate them, which means the honest position now is uncertainty rather than a preferred story.

And the mechanism split is more informative than the headline number. Thirteen percent, sixteen percent and twenty percent are all versions of one quantity. Whether automated and augmented occupations diverge is a different quantity and a better test, and it gets a fraction of the attention.

What it does not establish

That AI is not displacing entry-level workers. The automation-augmentation split is real and points that way, and the authors' narrowing moved the window rather than removing the finding.

That the monetary explanation is correct. It is a serious, well-argued alternative that predicts the observed pattern. It is not proven either.

That the data is unrepresentative. ADP payroll records covering millions of workers are among the best sources available for this question, and the dashboard team notes their sample complements rather than substitutes for nationally representative datasets.

And nothing about the eventual scale. This is an early-period measurement of a fast-moving change, and the honest range of futures it is consistent with is wide.

What is unresolved

Whether entry-level hiring recovers as rates ease. This is the natural experiment and it is in progress.

Whether the automation-augmentation gap widens. If it does, the AI account strengthens considerably.

What happens to the workers already displaced. A cohort that missed the bottom rung does not automatically join later, and there is little evidence on how such cohorts recover.

And whether occupational exposure measures are right. They are constructed from task descriptions and are proxies. If exposure is mismeasured, both the effect and its absence are mismeasured with it, which is a scope problem underneath both accounts.

The counter-argument

Emphasising the authors' caveat may overstate it. Researchers routinely report that results are sensitive to specification, and the broadest-control result is one specification among several. Treating it as a retraction reads more into a robustness note than the authors did, and the headline finding remains their published conclusion.

The monetary explanation has its own problem. It predicts entry-level weakness generally, and the observed weakness is concentrated in AI-exposed and specifically AI-automated occupations. EIG's account has to explain why the tightening cycle sorted itself by AI exposure, and the answer that exposed occupations are disproportionately in rate-sensitive sectors is plausible and not demonstrated.

Both may be right in proportions nobody can measure. The framing of competing explanations invites picking, when the likely truth is a mixture whose weights are not identifiable from the available data. This article's insistence on not choosing may itself understate how much of each is present.

And treating this as a measurement question can obscure the human one. A 22-year-old who cannot find a first job is in the same position whichever mechanism produced it, and precision about causation is worth less to them than it is to the argument.

The short version

ADP payroll records for millions of US workers show a 13% relative employment decline for ages 22 to 25 in the most AI-exposed occupations, 16% in a later version, and around 20% for young software developers since late 2022. Declines concentrate where AI automates rather than augments. Experienced workers are largely unaffected.

In February 2026 the same authors reported that with the broadest set of controls, the decline only becomes significant in 2024, and earlier declines were likely driven by non-AI factors. That narrowing was published by the researchers themselves and did not travel with their headline.

A competing account attributes the pattern to the sharpest monetary tightening in four decades, under which firms stop hiring at the bottom first, entry points vanish, and a disproportionate hit to 22 to 25 year olds is exactly what the theory predicts, with no reference to AI.

The two shocks share a start date, which makes them close to unidentifiable by controls alone. The AI account's best evidence is that automated and augmented occupations diverge, which a hiring freeze has no reason to produce. The macroeconomic account's best evidence is the age gradient and the 2024 significance boundary.

What both agree on is that entry-level employment in exposed occupations fell and young workers are bearing the adjustment. What separates them is a prediction: if it is monetary, the rung returns when hiring recovers. If it is automation, it does not. That test is running now.

Common questions

What did the study actually find? Using ADP payroll records covering millions of US workers through September 2025, Brynjolfsson, Chandar and Chen found a 13% relative employment decline for workers aged 22 to 25 in the most AI-exposed occupations after controlling for firm-level shocks, reported as 16% in a later version, with around a 20% decline for software developers aged 22 to 25 since late 2022. The declines concentrate in occupations where AI automates tasks rather than augmenting them, and employment for experienced workers largely held up.

What was the authors' later qualification? In a February 2026 update they reported that when the broadest set of controls is included, the timing of decline in AI-exposed occupations only becomes significant in 2024, and that earlier declines were likely influenced by non-AI factors. Generative AI reached wide adoption in late 2022, so this narrows the window in which the effect is attributable. It does not overturn the finding, and it was published by the same team in the same project.

Why does that matter? Because the 13% figure is quoted constantly and the qualification appears almost nowhere. A caveat issued by the authors themselves, published openly, still lost the race to their own headline. That is citation decay without a broken link or a decade of distance: the number was quotable and the qualification was not.

What is the competing explanation? That the pattern reflects the sharpest monetary policy tightening cycle in four decades rather than technological displacement. Rate rises from 2022 collapsed hiring, job postings fell, and firms that stop hiring stop at the bottom first because entry-level roles are marginal and experienced staff are retained. A disproportionate effect on 22 to 25 year olds is what that theory predicts, without reference to AI.

Why can't this be settled? Because both shocks began at the same time. Generative AI reached wide adoption in late 2022 and the tightening cycle ran from 2022, so any analysis must separate two large causes sharing a start date and acting on the same population. Controls alone cannot do it. Time can, which is why the natural test is what happens to entry-level hiring as monetary conditions ease.

Which evidence favours which side? The AI account's strongest asset is that declines concentrate in AI-automated rather than AI-augmented occupations, since a hiring freeze has no obvious reason to sort itself that way. The macroeconomic account's strongest assets are the age gradient, which a hiring freeze produces directly, and the authors' own finding that significance under the broadest controls begins in 2024 rather than at adoption.

What should someone watch next? Two things. Whether entry-level hiring recovers as rates ease, which is the natural experiment now running. And whether the gap between AI-automated and AI-augmented occupations widens as the macroeconomic shock recedes, which would strengthen the AI account considerably. The second is more informative than any further refinement of the headline percentage.

Does the uncertainty mean nothing is happening? No. Both accounts agree that entry-level employment in exposed occupations declined and that young workers are bearing the adjustment. The dispute is about mechanism, not about whether the pattern is real, and the mechanism matters mainly because it determines whether the effect unwinds when hiring recovers. For someone who cannot find a first job, the distinction is analytically important and practically cold.